This is wrong, a split collector PNP has partition noise. What you are saying boils down to taking two transistors biased with a battery or other filtered essentially noiseless voltage will have collector currents that are perfectly correlated this is not the case. Simply simulate it, a current mirror is a voltage (Vbe) controlled device, replace the Vbe with a battery the collector shot noise does not disappear..
Last edited:
You must have misinterpreted what I wrote. "Rather strong correlation" is not the same as "perfectly correlated". Of course they are not perfectly correlated!
I assume that the PNP mirror produces shot noise and that the shot noise of its output devices is independent. However, the mirror amplifies, so the shot noise of the input device and of the NPN that generates its input current dominates.
That is, it's fairly easy to see that n * sqrt(2 q I/n) > sqrt(2 q I) for n > 1, I don't need a simulator for that.
The amplified noise of the model NPN of the base current compensation circuit and of the input transistor of the PNP mirror ends up as correlated noise in the inputs.
I assume that the PNP mirror produces shot noise and that the shot noise of its output devices is independent. However, the mirror amplifies, so the shot noise of the input device and of the NPN that generates its input current dominates.
That is, it's fairly easy to see that n * sqrt(2 q I/n) > sqrt(2 q I) for n > 1, I don't need a simulator for that.
The amplified noise of the model NPN of the base current compensation circuit and of the input transistor of the PNP mirror ends up as correlated noise in the inputs.
Last edited:
By the way, for a given tail current of the NPN differential pair at the op-amp's input, the uncorrelated noise of the output transistors of the PNP current mirror is independent of the PNP current mirror's amplification factor n. That's because the output transistors of the PNP current mirror have to run at a collector current equal to the base current of the transistors of the NPN differential pair. Scale up n, and you have to scale down the current that goes into the PNP mirror, and you increase the correlated noise while keeping the uncorrelated part constant.
Does anyone understand why the PNP mirror has no emitter degeneration? Emitter degeneration resistors could have reduced its noise.
Does anyone understand why the PNP mirror has no emitter degeneration? Emitter degeneration resistors could have reduced its noise.
At these current levels they would need to be huge. There possibly is some confusion, I was talking more about partition noise in a split collector PNP not being accounted for. It was more than 20yr. ago when we investigated, I remember the improvement being less than expected.
A larger issue is that you would generally not use an amplifier in a region where the current noise is the major contributor. Take a MM phono stage for instance, how would you take advantage of this to make any net improvement in SNR?
OK, then why do you use split collector PNPs instead of separate PNPs, each with its own emitter resistor? At current levels around 4.5 uA, a resistance of 40 kohm should suffice.
The noise optimum is where collector shot noise and base shot noise (plus anything that's proportional to it) are equally strong. So indeed, current noise should not dominate, but it should produce almost half the total noise power. Of course only a part of that comes from the PNP mirror.
The noise optimum is where collector shot noise and base shot noise (plus anything that's proportional to it) are equally strong. So indeed, current noise should not dominate, but it should produce almost half the total noise power. Of course only a part of that comes from the PNP mirror.
Could be, I've never worked with split collector PNPs, they simply weren't available in the processes I worked in. What's the difference in noise compared to the paralleled unit PNP transistors that I'm used to?
Never had a definitive resolution, there is only one emitter so there was speculation on how shot noise would correlate at the collectors. There is also no resolution with respect to correlation on a JFET cascode with split drains.
In either case Ib comp makes the uncorrelated noise worse. I looked at some old computations and simulation agrees with what I measured 4.5uA is ~1.2pA noise so 4 totally uncorrelated currents are 2.4pA with a current mirror of individual devices this comes to 1.92pA or ~20% improvement. The argument started over the PNP's working some magic.
Last edited:
Also, a 40k resistor is huge in a standard process. Impossible to do with a base or emitter diffusion (huge area would be required) needs to be epi (essentially a JFET) or perhaps thin film.
That's probably very process-dependent. I've used pinched resistors in the hundreds of kilo-ohm range in the old days when I still worked in bipolar and BiCMOS processes, but none of those processes could cope with the supply voltages normally used for op-amps.
It can't have much impact; Hans's measurements of the LT1028 match with two to three digit accuracy to my calculations that treat a split-collector PNP as though it were a bunch of paralleled PNPs.
I have not looked at the 1028 datasheet in years, I forgot the graph that started the discussion. For some reason the noise at 1kHz line shows no change at all with the addition of the second 10k resistor this is clearly not possible.
Hans, Marcel sorry I missed this post I would have not made some earlier comments. All the numbers here are exactly as I remember.
To explain, the 797 uses a simple current mirror with gain of one so the unmatched number is root 2 times the basic Ib noise or ~1.6pA. If there was no correlation of any of the currents the matched and unmatched numbers would be the same. So there is partial correlation but we did not see the point in pushing it and did not spec matched noise.
The current mirror with gain makes the unmatched noise much worse so I guess it was a decision they had to make. If you are building something like a phono pre-amp matching source becomes very problematic
I have to take back the last comment the single ended noise is so much worse that the matched sources make a lot more difference.
Last edited:
Pinched resistors are what I called epi resistors; as i said, essentialy very long channel JFETs.
- Status
- Not open for further replies.